Monte Carlo — Stochastic Simulation for Risk

MONTE CARLO

Stochastic simulation for risk & reliability Sampling

Metropolis & Ulam (1949) · Manhattan project lineage · Distribution propagation

Monte Carlo methods estimate properties of an uncertain system by sampling its inputs from probability distributions, propagating each sample through the system model, and aggregating the outputs. Where analytical methods can give a single point estimate of a risk metric, Monte Carlo gives the full output distribution — so analysts can report not just expected loss but tail risk, percentiles, conditional value at risk, and sensitivity to each input.

Overview of the technique

Monte Carlo simulation was named for the casino district by Stanisław Ulam, John von Neumann and Nicholas Metropolis at Los Alamos in 1946–49 while developing weapons-physics calculations on early electronic computers. It rests on the law of large numbers: as the number of samples grows, the empirical distribution of model outputs converges to the true distribution. The method is general — it can attack any deterministic or stochastic computation that is too complex for analytical solution — and is now ubiquitous in physics, finance, project management, climate modelling, transport demand, reliability engineering, and probabilistic safety assessment.

A typical workflow: (1) define the system model Y = f(X₁, ..., Xₙ); (2) characterise each input by a probability distribution and capture dependencies (correlation matrices, copulas); (3) sample input vectors using pseudo-random or quasi-random sequences (Latin Hypercube, Sobol); (4) evaluate the model for every sample; (5) aggregate to produce the empirical distribution of Y, statistics (mean, variance, percentiles, exceedance probability) and sensitivity measures (Spearman rank, Sobol indices). Variance-reduction techniques — antithetic variates, control variates, importance sampling, stratified sampling — let you reach the necessary tail-precision with fewer samples. Markov-chain Monte Carlo (MCMC) extends the method to sampling from posterior distributions in Bayesian inference.

Figure 1. The Monte Carlo loop: inputs are sampled from their distributions, propagated through the model thousands of times, and aggregated into an output distribution with risk percentiles and sensitivities.

When to use it

Typical applications

Uncertainty propagation through complex risk models
Fault-tree quantification with uncertain basic-event data
Project schedule and cost risk (PERT, S-curves)
Trajectory and CFD-style stochastic simulation
Bayesian inference via MCMC sampling

Aviation relevance

Quantifies uncertainty in catastrophic failure rates (10⁻⁹/FH)
Underpins runway-incursion and TCAS encounter modelling
Used in UAS Specific Operations Risk Assessment (SORA) tail-risk
Drives ATM capacity and conflict-rate simulation in NAS analyses
Common in fatigue and crack-growth probabilistic modelling

Benefits

Full output distribution

Replaces a point estimate with an empirical distribution — letting analysts quote tail risk, exceedance probability and confidence bands rather than misleadingly precise means.

General-purpose

Works for any model that can be evaluated computationally; no closed-form analytical tractability required, so it scales to genuinely complex aerospace and ATM problems.

Sensitivity analysis built in

Sample-based sensitivity measures (Spearman, Sobol, regression-based) reveal which uncertain inputs dominate the output distribution and where data investment will pay off.

Couples to other methods

Pairs naturally with FTA, ETA, Bayesian networks and project-management models — a Monte Carlo wrapper around an existing model often unlocks honest uncertainty quantification.

Limitations

Garbage in, garbage out

Output distributions are only as good as the input distributions and dependence structure; defaulting to convenient (often Normal) priors can flatten genuine tail risk.

Slow convergence in tails

Standard sampling needs O(1/ε²) samples for relative precision ε — inefficient for very rare events; importance sampling or subset simulation are usually required for 10⁻⁶ – 10⁻⁹ regions.

Computationally expensive

For high-fidelity models (CFD, FEA, agent-based ATM) each evaluation may take minutes; surrogate models, polynomial chaos and emulators are often necessary to make Monte Carlo tractable.

Reproducibility & auditing

Pseudo-random seeds, software versions and sampling schemes must be archived; otherwise certification reviewers cannot verify the run, and minor implementation changes can shift tail estimates noticeably.

In short

Monte Carlo turns any deterministic risk model into a probability machine: sample the inputs, run the model, and read off percentiles and sensitivities. The discipline is in the input distributions, the dependence structure, and choosing variance-reduction methods that reach the tail you actually care about.

References (APA 7)

Metropolis, N., & Ulam, S. (1949). The Monte Carlo method. Journal of the American Statistical Association, 44(247), 335–341.

Robert, C. P., & Casella, G. (2004). Monte Carlo statistical methods (2nd ed.). Springer.

Rubinstein, R. Y., & Kroese, D. P. (2017). Simulation and the Monte Carlo method (3rd ed.). Wiley.

Vose, D. (2008). Risk analysis: A quantitative guide (3rd ed.). Wiley.

Helton, J. C., & Davis, F. J. (2003). Latin hypercube sampling and the propagation of uncertainty in analyses of complex systems. Reliability Engineering & System Safety, 81(1), 23–69.

International Electrotechnical Commission. (2019). Risk management — Risk assessment techniques (IEC 31010:2019). IEC.